AI in Education
AI in education refers to the use of machine learning and language models to teach, assess and support learners and teachers: tutoring dialogue, adaptive practice, automated feedback, lesson and quiz generation, and administrative help. The field is older than chat assistants, and it is organised around a single recurring question: can software supply some of the benefit of one-to-one human tutoring at the scale of a classroom or a country?
The motivating target is Bloom's two-sigma problem. In 1984 the educational psychologist Benjamin Bloom reported in Educational Researcher that students who received one-to-one tutoring with mastery learning performed about two standard deviations better than students in conventional classes, placing the average tutored student above roughly 98 percent of the classroom group. He called finding group methods that could match this the "2 sigma problem." The result came from a few small studies by Bloom's doctoral students, and later reviews find smaller effects: VanLehn's 2011 review estimated human tutoring at around 0.79 standard deviations and well-built intelligent tutoring systems at around 0.76. Even so, the large gap between tutoring and classroom instruction continues to motivate the field.
Intelligent tutoring systems made the first serious attempt. Jaime Carbonell's SCHOLAR (1970) held tutorial dialogue about geography. John Anderson's group at Carnegie Mellon built cognitive tutors for algebra and programming, grounded in a model of the skills students must acquire, and these became the commercial Carnegie Learning products used in US schools. They track mastery skill by skill, give step-level hints and choose the next problem. Adaptive platforms such as Duolingo apply similar ideas to practice sequencing. The strength of these systems is a well-specified domain where every step can be checked; the weakness is the cost of authoring that structure for each subject.
Language models offer breadth in exchange for reliability. Khan Academy introduced Khanmigo in 2023, built on OpenAI's GPT-4, as a tutor and teaching assistant that is instructed to ask guiding questions instead of handing over answers. Duolingo Max and other products took similar steps the same year. Models can explain a concept several ways, role-play a historical figure, or give feedback on an essay draft at any hour. They can also make arithmetic and factual mistakes while sounding certain, and a student who accepts a wrong explanation has no easy way to notice. Strong designs constrain the model with curated content, verify calculations with tools and keep teachers informed.
Cheating and integrity are the most visible conflict. After ChatGPT appeared in November 2022, students could generate essays in seconds. New York City's education department blocked the tool on school networks in January 2023 and lifted the restriction months later. Detector software was offered in response, including Turnitin's AI-writing indicator in 2023, but detectors give probabilistic scores and make false accusations; a 2023 study by Liang and colleagues in Patterns found that several detectors misclassified writing by non-native English speakers as machine-generated at high rates. Many institutions have moved toward redesigned assessment: in-class writing, oral defenses, drafts and process records, and explicit policies on permitted use.
Pitfalls and misconceptions. The deepest risk is not cheating but substitution: a tool that does the thinking can raise completed assignments while lowering learning, because understanding requires effortful practice. Equity matters, since access, language support and teacher training vary. Student data is sensitive, and laws such as FERPA and COPPA in the United States constrain how it is handled. Personalized does not always mean effective; evidence for many commercial claims is thin. Related entries cover personalization and conversational AI, while this one concerns the pedagogy that decides whether those capabilities help a student learn.
Read AI in education as the old tutoring ambition meeting a new, fluent but fallible tool: it helps when it keeps students doing the thinking and fails when it does the thinking for them.